Skip to content

feat: Added support for aten::tile converter #2105

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Merged
merged 2 commits into from
Aug 5, 2023
Merged

feat: Added support for aten::tile converter #2105

Show file tree
Hide file tree
Changes from 1 commit
Commits
Show all changes
2 commits
Select commit Hold shift + click to select a range
6b51057
feat: Added support for aten::tile converter
Jul 11, 2023
b7b2725
feat: Moved from converter to lowering pass
Aug 3, 2023
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
115 changes: 61 additions & 54 deletions core/conversion/converters/impl/expand.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -194,6 +194,60 @@ bool add_expand_dynamic(
return true;
}

bool add_repeat(ConversionCtx* ctx, const torch::jit::Node* n, args& args, const std::string& layer) {
auto in = args[0].ITensorOrFreeze(ctx);
auto input_dims = in->getDimensions();
auto repeats = args[1].unwrapToIntList().vec();
int repeats_rank = repeats.size();
TORCHTRT_CHECK(
repeats_rank >= input_dims.nbDims,
"Number of repeat dimensions cannot be smaller than number of input dimensions");

auto num_expand_dims = repeats_rank - input_dims.nbDims;

if (ctx->input_is_dynamic) {
int input_rank = input_dims.nbDims;
int output_rank = repeats_rank;
auto new_input_shape_tensor = concat(output_rank, input_rank, ctx, in);

auto shuffle = ctx->net->addShuffle(*in);
shuffle->setInput(1, *new_input_shape_tensor);
in = shuffle->getOutput(0);
} else {
if (num_expand_dims > 0) {
nvinfer1::Dims reshape_dims;
reshape_dims.nbDims = repeats.size();
for (int i = 0; i < num_expand_dims; i++) {
reshape_dims.d[i] = 1;
}
for (int i = 0; i < input_dims.nbDims; i++) {
reshape_dims.d[num_expand_dims + i] = input_dims.d[i];
}
// Add a reshape layer to expand dims
auto reshape_layer = ctx->net->addShuffle(*in);
reshape_layer->setReshapeDimensions(reshape_dims);
in = reshape_layer->getOutput(0);
LOG_DEBUG("Input reshaped to : " << in->getDimensions() << " from " << input_dims);
}
LOG_DEBUG("Repeats: " << repeats);
}

// Concat across all repeat axes.
for (int i = repeats.size() - 1; i >= 0; --i) {
std::vector<nvinfer1::ITensor*> tensors_vec;
for (int j = 0; j < repeats[i]; j++) {
tensors_vec.push_back(in);
}
auto concat_layer = ctx->net->addConcatenation(tensors_vec.data(), tensors_vec.size());
concat_layer->setAxis(i);
in = concat_layer->getOutput(0);
}

auto out = ctx->AssociateValueAndTensor(n->outputs()[0], in);
LOG_DEBUG(layer << " layer output tensor shape: " << out->getDimensions());
return true;
}

auto expand_registrations TORCHTRT_UNUSED =
RegisterNodeConversionPatterns()
.pattern(
Expand Down Expand Up @@ -230,59 +284,7 @@ auto expand_registrations TORCHTRT_UNUSED =
.pattern(
{"aten::repeat(Tensor self, int[] repeats) -> (Tensor)",
[](ConversionCtx* ctx, const torch::jit::Node* n, args& args) -> bool {
auto in = args[0].ITensorOrFreeze(ctx);
auto input_dims = in->getDimensions();
auto repeats = args[1].unwrapToIntList().vec();
int repeats_rank = repeats.size();
TORCHTRT_CHECK(
repeats_rank >= input_dims.nbDims,
"Number of repeat dimensions cannot be smaller than number of input dimensions");
auto num_expand_dims = repeats_rank - input_dims.nbDims;

if (ctx->input_is_dynamic) {
int input_rank = input_dims.nbDims;
int output_rank = repeats_rank;
auto new_input_shape_tensor = concat(output_rank, input_rank, ctx, in);

// Add a reshape layer to expand dims
auto shuffle = ctx->net->addShuffle(*in);
shuffle->setInput(1, *new_input_shape_tensor);
in = shuffle->getOutput(0);
} else {
if (num_expand_dims > 0) {
nvinfer1::Dims reshape_dims;
reshape_dims.nbDims = repeats.size();
for (int i = 0; i < num_expand_dims; i++) {
reshape_dims.d[i] = 1;
}
for (int i = 0; i < input_dims.nbDims; i++) {
reshape_dims.d[num_expand_dims + i] = input_dims.d[i];
}
// Add a reshape layer to expand dims
auto reshape_layer = ctx->net->addShuffle(*in);
reshape_layer->setReshapeDimensions(reshape_dims);
in = reshape_layer->getOutput(0);
LOG_DEBUG("Input reshaped to : " << in->getDimensions() << " from " << input_dims);
}
LOG_DEBUG("Repeats: " << repeats);
}

// Concat across all repeat axes.
// TODO: Implementation might not be performant. Explore other strategies to improve performance.
for (int i = repeats.size() - 1; i >= 0; --i) {
std::vector<nvinfer1::ITensor*> tensors_vec;
for (int j = 0; j < repeats[i]; j++) {
tensors_vec.push_back(in);
}
auto concat_layer = ctx->net->addConcatenation(tensors_vec.data(), tensors_vec.size());
concat_layer->setAxis(i);
in = concat_layer->getOutput(0);
}

auto out = ctx->AssociateValueAndTensor(n->outputs()[0], in);

LOG_DEBUG("Repeat layer output tensor shape: " << out->getDimensions());
return true;
return add_repeat(ctx, n, args, "Repeat");
}})
.pattern(
{"aten::repeat_interleave.self_int(Tensor self, int repeats, int? dim=None, *, int? output_size=None) -> (Tensor)",
Expand Down Expand Up @@ -395,6 +397,11 @@ auto expand_registrations TORCHTRT_UNUSED =

return true;
}})
.pattern(
{"aten::tile(Tensor self, int[] dims) -> (Tensor)",
[](ConversionCtx* ctx, const torch::jit::Node* n, args& args) -> bool {
return add_repeat(ctx, n, args, "Tile");
}})
.pattern(
{"aten::meshgrid(Tensor[] tensors) -> (Tensor[])",
[](ConversionCtx* ctx, const torch::jit::Node* n, args& args) -> bool {
Expand Down Expand Up @@ -484,4 +491,4 @@ auto expand_registrations TORCHTRT_UNUSED =
} // namespace converters
} // namespace conversion
} // namespace core
} // namespace torch_tensorrt
} // namespace torch_tensorrt
120 changes: 120 additions & 0 deletions tests/core/conversion/converters/test_expand.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -670,6 +670,126 @@ TEST(Converters, ATenRepeatInterleave3dScalarNoDimConvertsCorrectlyWithDynamicIn
ASSERT_TRUE(torch_tensorrt::tests::util::almostEqual(jit_results[0], trt, 2e-6));
}

TEST(Converters, ATenTileConvertsCorrectly) {
const auto graph = R"IR(
graph(%x.1 : Tensor):
%2 : int[] = prim::Constant[value=[4, 1]]()
%3 : Tensor = aten::tile(%x.1, %2)
return (%3))IR";

auto g = std::make_shared<torch::jit::Graph>();

torch::jit::parseIR(graph, g.get());

auto in = at::randint(1, 10, {1, 3}, {at::kCUDA});

auto jit_in = at::clone(in);
auto params = torch_tensorrt::core::ir::get_static_params(g->inputs(), {});
auto jit_results = torch_tensorrt::tests::util::RunGraph(g, params, {jit_in});

auto trt_in = at::clone(jit_in);
params = torch_tensorrt::core::ir::get_static_params(g->inputs(), {});
auto trt_results = torch_tensorrt::tests::util::RunGraphEngine(g, params, {trt_in});

ASSERT_TRUE(torch_tensorrt::tests::util::almostEqual(jit_results[0], trt_results[0], 2e-6));
}

TEST(Converters, ATenTileRepeatRankConvertsCorrectly) {
const auto graph = R"IR(
graph(%x.1 : Tensor):
%2 : int[] = prim::Constant[value=[4, 1, 2]]()
%3 : Tensor = aten::tile(%x.1, %2)
return (%3))IR";

auto g = std::make_shared<torch::jit::Graph>();

torch::jit::parseIR(graph, g.get());

auto in = at::randint(1, 10, {1, 3}, {at::kCUDA});

auto jit_in = at::clone(in);
auto params = torch_tensorrt::core::ir::get_static_params(g->inputs(), {});
auto jit_results = torch_tensorrt::tests::util::RunGraph(g, params, {jit_in});

auto trt_in = at::clone(jit_in);
params = torch_tensorrt::core::ir::get_static_params(g->inputs(), {});
auto trt_results = torch_tensorrt::tests::util::RunGraphEngine(g, params, {trt_in});

ASSERT_TRUE(torch_tensorrt::tests::util::almostEqual(jit_results[0], trt_results[0], 2e-6));
}

TEST(Converters, ATenTileConvertsCorrectlyWithDynamicInput) {
const auto graph = R"IR(
graph(%x.1 : Tensor):
%2 : int[] = prim::Constant[value=[4, 1]]()
%3 : Tensor = aten::tile(%x.1, %2)
return (%3))IR";

auto g = std::make_shared<torch::jit::Graph>();

torch::jit::parseIR(graph, g.get());

auto in = at::randint(1, 10, {1, 3}, {at::kCUDA});

auto jit_in = at::clone(in);
auto params = torch_tensorrt::core::ir::get_static_params(g->inputs(), {});
auto jit_results = torch_tensorrt::tests::util::RunGraph(g, params, {jit_in});

auto trt_in = at::clone(jit_in);
params = torch_tensorrt::core::ir::get_static_params(g->inputs(), {});
auto trt_results = torch_tensorrt::tests::util::RunGraphEngineDynamic(g, params, {trt_in});

ASSERT_TRUE(torch_tensorrt::tests::util::almostEqual(jit_results[0], trt_results[0], 2e-6));
}

TEST(Converters, ATenTile3dConvertsCorrectly) {
const auto graph = R"IR(
graph(%x.1 : Tensor):
%2 : int[] = prim::Constant[value=[2, 2, 2]]()
%3 : Tensor = aten::tile(%x.1, %2)
return (%3))IR";

auto g = std::make_shared<torch::jit::Graph>();

torch::jit::parseIR(graph, g.get());

auto in = at::randint(1, 10, {2, 3, 2}, {at::kCUDA});

auto jit_in = at::clone(in);
auto params = torch_tensorrt::core::ir::get_static_params(g->inputs(), {});
auto jit_results = torch_tensorrt::tests::util::RunGraph(g, params, {jit_in});

auto trt_in = at::clone(jit_in);
params = torch_tensorrt::core::ir::get_static_params(g->inputs(), {});
auto trt_results = torch_tensorrt::tests::util::RunGraphEngine(g, params, {trt_in});

ASSERT_TRUE(torch_tensorrt::tests::util::almostEqual(jit_results[0], trt_results[0], 2e-6));
}

TEST(Converters, ATenTile3dConvertsCorrectlyWithDynamicInput) {
const auto graph = R"IR(
graph(%x.1 : Tensor):
%2 : int[] = prim::Constant[value=[2, 2, 2]]()
%3 : Tensor = aten::tile(%x.1, %2)
return (%3))IR";

auto g = std::make_shared<torch::jit::Graph>();

torch::jit::parseIR(graph, g.get());

auto in = at::randint(1, 10, {2, 3, 2}, {at::kCUDA});

auto jit_in = at::clone(in);
auto params = torch_tensorrt::core::ir::get_static_params(g->inputs(), {});
auto jit_results = torch_tensorrt::tests::util::RunGraph(g, params, {jit_in});

auto trt_in = at::clone(jit_in);
params = torch_tensorrt::core::ir::get_static_params(g->inputs(), {});
auto trt_results = torch_tensorrt::tests::util::RunGraphEngineDynamic(g, params, {trt_in});

ASSERT_TRUE(torch_tensorrt::tests::util::almostEqual(jit_results[0], trt_results[0], 2e-6));
}

TEST(Converters, ATenMeshGridConvertsCorrectly) {
const auto graph = R"IR(
graph(%x : Tensor, %y : Tensor, %z : Tensor):
Expand Down